The Classification Of ECG Signals Based On Improved U-Net++ Networks

Cardiovascular disease patients often have arrhythmias before the onset of cardiac arrhythmias,most of the diagnosis of arrhythmias based on the results of electrocardiogram.Electrocardiogram(ECG)can directly reflect the patient’s heart health,so it is widely used in the diagnosis of various cardiovascular diseases.The automatic recognition algorithm of ECG signals,as an auxiliary medical means,has high recognition accuracy and high clinical application value,which can reduce the pressure of experts and doctors.Therefore,this thesis constructs an automatic recognition and diagnosis model of ECG signals based on deep learning to complete the classification of various types of ECG signals.The main research contents of this thesis are as follows:1.The MIT-BIH arrhythmia database is the most widely used in current research ECG database,this thesis chose the largest number of four different types of sample point length of1200 segments of ECG signal timing records as the experimental data,but some of the exception class data quantity is less,as compared to normal classes such as ECG types according to this data imbalance problem,this thesis uses the improvement the depth of the convolution generated against network URNET-GAN to complete some type of ECG signal data expansion experiment;U-NET++ structure is used as the generator part of URNET-GAN,and the residual network is used as the discriminator part of URNET-GAN.Be due to the experiment data compared with real ECG signal recording is not smooth,in this thesis,based on the wavelet threshold denoising method,first by using wavelet,wavelet function promote reusing wavelet decomposition to spread to all the signal data in various scales,through the study of the threshold value of high frequency coefficients of each scale set,for each dimension of noise denoising,finally through the wavelet inverse transformation through wavelet reconstruction of each layer in turn end up with a smooth expansion of data.2.In the case that the original ECG signal fragments are not preprocessed and the generated extended data is preprocessed by wavelet threshold filtering to remove noise,the two parts of data are integrated to complete the division of the training set,the verification set and the test set.In this thesis,the improved U-NET++ network was used to identify and classify four different types of ECG data: normal,premature ventricular contraction,left bundle branch block and right bundle branch block.The 5-layer one-dimensional U-NET++ network model with the convolution kernel size of 32 adopted in this thesis finally obtains the overall accuracy of 98.10% on the training set,and has good results on the precision rate,recall rate,F1 value and other indexes of the test set.Under the same experimental parameters,compared with the undersampled data,the accuracy of the experimental model of the expanded data set with URNet-GAN data is improved by 0.85%.Under the same experimental data set,the accuracy is 1.05% higher than that of U-NET model.